The Impact of Job Mapping on Random Network Topology

摘要

A number of small parallel applications run on datacenters and supercomputers simultaneously. Job mapping becomes crucial to improving system utilization and application execution. Fragmentation of unused compute nodes could not be assigned for an incoming job since it may largely harm communication abilities between non-adjacent compute nodes. In this case, however, incoming jobs are likely to be pending on the overloaded system because they have to wait for the release of adjacent occupied compute nodes. In this study, we explore job mapping on random topology for the purpose of improving job scheduling ability. Ideally, a diverse application workload can be better supported disregarding its interconnection network topology with a certain time-space tradeoff. Our simulation results demonstrate that, over 3-D torus interconnection networks, the embedding of random topology performs better than that of 2-D mesh by 84% and seems comparable to that of 3-D mesh in terms of job scheduling performance. Over random topologies, the scheduling performance can be much improved by the embedding of random topologies especially for dealing with dozens of intensively incoming jobs. Overall, job mapping on random guest topology over random host topology presents the best job scheduling performance among all the cases in our evaluation.